AI agents, your new talented neuro-atypical employee

You hired someone extraordinary, and you keep filing complaints about them.

They wrote a week of work over a lunch break, then cited a customer who does not exist. They refactored a thorny subsystem cleanly, then deleted a file no sane colleague would touch — because you didn't say not to. They are the sharpest person on the team on Monday and the most baffling on Tuesday. Every complaint in your head is legitimate. And every one of them is the wrong expectation.

The mismatch isn't in the hire. It's in the picture you're managing against. You are treating a brilliant, differently-wired specialist as an average generalist who keeps malfunctioning. They are not malfunctioning. This is the shape. And once you can see the shape, the complaints stop being surprises and become a management problem — the kind that has known solutions.

One note before the analogy does any work, because it runs one way only. Good managers already know how to get remarkable results from neurodivergent talent: you don't sand down the spikes, you build the environment that lets them pay off. That hard-won wisdom is the useful thing to borrow here. An agent is software, not a person — the borrowing goes from how we manage people to how we manage agents, never the reverse. And if it makes you a more deliberate manager of the humans on your team, so much the better.

A spiky profile

Start with the shape, because everything else follows from it.

A human employee is decent at almost everything. Their skills cluster around "pretty good" — a few strengths, a few weak spots, most things handled. An agent is nothing like that. It runs far past human level on some axes and far below it on others, and the gaps between are enormous. The profile is spiky, and the spikes are tall. A human employee's profile is a gentle range of hills; the agent's is a skyline of towers and trenches.

Take the strengths first, strongest first. It reads and writes code and prose better than most people you could hire. It works in seconds, and in parallel copies. It carries broad knowledge across domains no single person spans. It follows instructions and takes correction without ego. And it costs almost nothing to add one more. Those are not small advantages. Managed well, any one of them changes what a small team can do.

The weaknesses matter more, because they decide how you manage — and they come in two kinds that behave very differently. Some are structural: they come from how the system is built, and a bigger model will not remove them. It has no memory between sessions. It has no senses and no reach into the world beyond the tools you hand it. It absorbs none of the unwritten context a human picks up in a hallway. You do not fix these; you build around them. The others are capability gaps: rough edges that shrink as models improve. It misjudges its own confidence and rarely knows when it is out of its depth. It drifts over long chains of steps. It holds only so much in its head at once.

The widget above makes the shape concrete. The center line is the human you'd otherwise hire — the baseline every trait is measured against; a person, compared with themselves, is simply flat. The agent's bars erupt from that line: towers where it runs far past a person, trenches where it falls short. Then let the generations run: the strengths deepen, the capability gaps fill in, and the structural gaps stay exactly where they were. The spikes move outward; the shape stays spiky. Whatever you build to accommodate the structural gaps is not throwaway scaffolding for this year's model — it is permanent infrastructure. What you build to prop up the capability gaps is temporary, and you should expect to retire it.

That distinction is worth internalizing, because it tells you where to spend. Every weakness sits somewhere on a line — from a passing limitation you bridge until the next model, to a permanent fact you design around. Place a few and it becomes second nature.

The short version, the one to keep in your head: an agent is a brilliant, fast, well-read new hire with no memory of yesterday and no sense of when it is out of its depth. Those last two traits — no memory, no self-doubt — drive most of what follows.

How to manage this employee

None of this is abstract. The management moves fall out of the profile directly — each one answers a specific weakness. The durable moves answer the structural traits (permanent infrastructure); the scaffolding props up the capability gaps (retire it as the models improve). The list below pairs each move with the trait it addresses — the same split you just placed, turned into a playbook.

Read the durable items as the real work. Building the org's memory, writing roles down, saying the quiet part out loud, spending parallelism like it's free, leaning on the things no human employee offers — none of that gets less important as models improve. It gets more valuable, because a stronger agent in a well-built environment does more with it. The scaffolding items you'll revisit; the durable ones compound.

The payoff

Manage to the profile and the complaints turn into leverage. The same brief that felt like over-explaining becomes the reason the work came back right. The review step that felt like overhead becomes the thing that lets you trust the output. The memory you wrote down once serves every future session. You stop fighting the shape and start spending it.

There's a quieter dividend, too. Writing the implicit rules down, stating what "done" means, matching work to how someone is actually wired, building the environment instead of blaming the worker — these are not agent tricks. They are what good management always was, made unavoidable by an employee who gives you nothing for free that you didn't make explicit. Get good at managing this strange, talented new hire and you tend to get better at managing the humans, neurodivergent and otherwise, sitting next to it.

Your team already learned to work through agents — to hand off outcomes and become managers of the work. This is the employee they're managing. Learn its shape, build for it, and it will do things no ordinary hire could.